A self-supervised depth-aware method for pose optimization in hybrid bronchoscopic navigation

Area of Science:

• Medical Imaging
• Computer Vision
• Robotics

Background:

• Pose optimization is critical for hybrid vision-electromagnetic navigation systems, aligning bronchoscopic views with CT scans.
• Discrepancies between virtual (CT) and real bronchoscopic scenes cause appearance and scale ambiguities, hindering accurate pose estimation.
• Existing methods struggle with the modality gap, impacting the reliability of bronchoscope positioning.

Purpose of the Study:

• To present a self-supervised pose estimation method, Depth-aware Attention Pose Network (DaAPNet), to address ambiguities in hybrid navigation systems.
• To improve the accuracy and robustness of aligning virtual bronchoscopic views with real-time data.
• To mitigate appearance and scale ambiguities inherent in cross-modality pose estimation.

Main Methods:

• Developed DaAPNet, integrating a Depth-aware Attention module to leverage depth information for resolving structural ambiguities.
• Implemented an entropy-based uncertainty-weighted masking mechanism to suppress unreliable appearance cues.
• Incorporated a scale prediction module and pose consistency loss to enforce geometric consistency and temporal stability.

Main Results:

• DaAPNet demonstrated improved performance on virtual and phantom bronchus datasets.
• Qualitative results showed effective mitigation of appearance ambiguity by accurately representing bronchial bifurcations.
• Quantitative analysis revealed restoration of scale consistency, reducing translation error by 35.99%.

Conclusions:

• DaAPNet offers a robust solution for self-supervised pose estimation in hybrid bronchoscopic navigation.
• The method effectively addresses cross-modality ambiguities, enhancing the precision of bronchoscope alignment.
• This advancement contributes to more reliable and accurate image-guided interventions in pulmonology.